Thoughts about a WebGL-Next

Datetime:2016-08-22 23:09:50          Topic: WebGL           Share

TL;DR: what a WebGL2 successor could look like and why it can’t be ‘WebVulkan’

The topic of ‘WebVulkan’ as a WebGL2 successor is coming up from time to time, here’s my 2ct since I also spend a lot of time thinking about what a modern 3D-API ‘for the rest of us’ should look like.

For the record: I’m not associated with any browser, GPU or 3D-API maker, so this is purely from the ‘receiving end’ perspective ;)

Here are the current drawbacks of WebGLwhich should be fixed in a hypothetical ‘WebGL-Next’ (all IMHO):

The WebGL API is too granular

Just like GL, the WebGL API is too granular, too many calls are required per frame to get something interesting on screen. In a native GL implementation this isn’t a big problem, it only becomes a problem when running in JS. GL calls are cheap in a native environment, but much more expensive in a JS engine, especially if JS objects are involved that might put pressure on the garbage collector.

In WebGL1, the main problem areas are: render state updates, vertex attribute definition, uniform updates. WebGL2 has vertex array objects and uniform blocks which improve 2 of those. The granular render state update calls remain a problem though.

Because of the unfortunate combination of a granular API and a high call-overhead the same batching tricks like back in the D3D9 days must be performed when using WebGL. A lot of code complexity is needed to avoid redundant calls into WebGL, and even then the overhead is much higher compared to talking to a good native GL driver.

The overhead also exists because WebGL needs to peform some additional security-related checks (e.g. validate that vertex indexes are not out-of-bounds, which makes dynamic buffer updates more expensive). Also, on Windows, WebGL usually runs on top D3D9 or D3D11 instead of GL, which may add additional overhead.

This results in very different performance characteristics between WebGL and native GL implementations, code that performs well on OpenGL or GLES2 may perform badly on WebGL. On the other hand, WebGL is secure and extremely well tested for GLES2 conformance. The only 3D-API which comes even close to ‘write once run everywhere’.

From my experience, WebGL on desktop browsers can do about 5000 traditional draw calls at 60Hz (1 uniform update + 1 draw call), the differences between browser, underlying GPU or driver don’t matter that much. WebGL code will be much sooner CPU-bound than GPU bound, even on low-end GPUs, it is important to keep that in mind when writing WebGL code.

For comparison, a good OpenGL driver on Windows reaches over 100k draw calls before dropping below 60Hz (but at the same time a bad GL driver on an integrated GPU only reaches slightly above 10k).

WebGL can’t directly access GPU memory

This is probably the most critical limitation for a move to a more modern 3D API, but it exists for a good reason: Javascript is not allowed to directly write to GPU-accessible memory because: security.

I am not that much into browser security tech to know whether this limitation is forever cast in stone or whether there are watertight secure ways to write GPU memory directly from JS, but I think that would be a very tough nut to crack for a ‘WebGL next’. On the other hand, NaCl also did the ‘impossible’ to run actual native code securely in the browser, so may be there’s a clever way around.

The problem is that the modern 3D APIs are pretty much all about directly sharing memory between the CPU and GPU, and they don’t give a shit about whether the user-code does something wrong, instead the GPU will simply go belly up and sometimes take the whole system with it. Which is not really acceptable on the web of course.

WebGL can’t be efficiently threaded

There have been several attempts to drive WebGL from worker threads, but just as the failed attempts in traditional native 3D APIs, not much came of it. The modern 3D APIs have arrived at command lists to solve the threading problem in a simple and straight-forward way. Instead of directly feeding the GPU from different threads, command-lists are built on CPU-threads completely independent from each other, and then moved over to the main thread where they are queued for execution on the GPU.

WebGL has no shader byte code

This is a relatively small problem compared to the others, but each browser vendor has to implement its own GLSL compiler in the browser. And while WebGL is very well conformance-tested, sometimes a shader compiles fine in one browser but not in another.

Why a ‘WebVulkan’ makes no sense

Again all IMHO:

  • obviously: making GPU memory directly accessible from JS would be a tough challenge to harden for security, but Vulkan is all about giving the CPU direct access to GPU resources
  • Vulkan is not so much a 3D-API as a meta-API to build your own 3D-API. It no longer even pretends to be usable without a wrapper consisting of thousands of lines of boilerplate code. In that sense it is more an ‘anti-API’. I haven’t made up my mind yet whether this is good or bad. What I know is that writing Vulkan code feels like ‘work’, there’s not a lot of fun in it. It just gives you a box of screws and bolts and says: here’s everything you need to build your own flavour of D3D or GL. Go ahead and try to do better. D3D12 is nearly the same. It is wasted effort to bring this mindset to the web , especially since the performance advantages would be lost in the noise (even on native platforms, programmers struggle to beat D3D11 with their own specialized rendering code, writing Vulkan or D3D12 code that performs well across all GPU vendors is a really damn hard problem. Only Metal tries to offer a 3D API that’s usable by mere humans without a sanity layer inbetween.
  • None of the modern 3D APIs are available on all platforms:
    • Vulkan is Windows/Linux/Android only
    • D3D12 is Windows10/UWP only
    • Metal is iOS/OSX only
  • Trying to emulate Vulkan on D3D12 or even Metal (similar to what ANGLE does with GLES2-on-D3D9) is a wasted effort. All the performance advantages would be lost, in exchange for an over-complicated API.

Why a modern Web-3D-API still makes sense

Aside from the hideously complex manual resource management in Vulkan and D3D12 the new APIs also have good parts that would help to fix WebGL’s shortcomings, and even make it a simpler 3D-API than it is today:

  • Pipeline State Objects: these are opaque objects which bundle all granular render state, vertex layout definition and shaders into a single, immutable object. This would easily be the one change with the most benefit for WebGL. Instead of dozens of granular calls, a single call would reconfigure the entire rendering pipeline. The user-code would no longer need to implement state caching in order to reduce calls into WebGL, and a whole class of bugs would simply vanish where the render pipeline configuration is left in a inconsistent state because of a few forgotten render state updates.

  • Command Lists: These would solve the whole ‘rendering from web workers’ in a very simple and elegant way. The worker thread only needs to know opaque resource handles (should it also be able to create resources? hmm…). Render commands are recorded into command lists on worker threads, and than moved over to the main thread where the command lists are enqueued for execution on the GPU.

  • Render Passes: Vulkan- or Metal-style render passes enable an important class of optimizations for tiled-renderers, where multiple render passes can happen per-tile on the GPU without having to store pass-results in video memory.

  • SPIR-V shader byte code: SPIR-V is easier to evaluate and cross-translate than GLSL. Shader language frontends would still make sense, these could be implemented as JS modules instead of baking them into the browser, but the ‘web guys’ are now also getting used to a separate compile step (e.g. Typescript), so I don’t see a problem with compiling shader code before deployment.

Resource Management (buffers & images)

Here I would vote for ease-of-use instead of explicit control, the same way that D3D11 and especially Metal does it.

Instead of explicitely managing all resource memory and resource state transitions, the application code provides an intent how it is going to use a resource, and the details are handled inside the 3D API just like in D3D11 or (with slightly more control) in Metal.

It would be nice however to still get rid of redundant memory copies or granular API calls. For instance there should be a ‘buffer mapping’ where the JS code can directly write vertex/index/uniform updates, and then may be call a ‘didModifyRange’ similar to Metal on OSX to inform the 3D API about dirty regions. Under the hood, the browser would still need to validate the modified data and very likely do a separate copy to GPU memory though.

It should be possible to record all dynamic shader uniform updates for an entire frame into one big buffer without having to explicitely request access to the buffer more than once (at the start of the frame), for each draw call, record a simple buffer offset to tell the next draw call where its uniform data starts, and only do a single ‘didModifyRange’ for the entire buffer right before enqueueing the command list. This is possible in Metal, but not in D3D11 until D3D11.3 (if I remember right).

There’s one important difference between D3D11 and Metal regarding dynamic buffer updates: D3D11 has internal buffer renaming to make sure that the CPU doesn’t overwrite areas that the GPU might currently read. In Metal the application code needs to take care of this. The simplest solution in Metal is to implement a per-frame double-buffering scheme, and Metal is open enough to allow more elaborate schemes.

The End

So that’s it. My highly inofficial proposal for the next WebGL ;) The big difference to the ‘WebVulkan’ approach is that it focuses on a very simple API which would provide room for performance improvements over WebGL without the brute-force approach of D3D12 or Vulkan, and it could even be implemented on top of GLES2 if needed.

The days of the ‘One 3D-API To Rule Them All’ are over, arguably this was never true since OpenGL implementations differed so much in important details that each GL implementation was its own little platform with different feature sets and performance behaviour. WebGL is a bit of a ‘unicorn’ because it did the right thing at the right time, and in a very pragmatic way.

Being very similar to GL/GLES2 definitely drove WebGL’s success. But Vulkan and D3D12 have a very different purpose than GLES2. They are not meant for consumption ‘by the rest of us’, but written for a small elite of highly specialized rendering engineers. In my opinion this is a dangerous road never taken by any (surviving) 3D-API. Up until D3D11, each D3D API became easier to use, performance improvements were a nice side effect. In this regard, OpenGL was always a lost cause because it simply packed new stuff on old stuff, resulting in a terrible mess.

When I look at the current situation, I feel thrown back to around 1997, when each GPU vendor created its own shitty little 3D API, and Microsoft created that terrible mess that was Direct3D3 (when I wrote the D3D12 backend for Oryol, long buried painful memories of D3D3’s execute buffers surfaced more than once).

The whole situation would be different if Metal would be a child of Khronos with a plain C-API and supported by all platform owners. In that case the situation would be crystal clear: let’s create a WebMetal that maps 1:1 to this hypothetical ‘Khronos Metal’, and may be in some alternate reality this is even happening. But alas we’re stuck in this shitty universe and need to come up with another solution ;)

PS: if you want to get a better idea of what such a highly simplified 3D-API could look like, have a look here at the Oryol Gfx interface . Granted, this leaves out a lot of modern stuff since it needs to map back to GLES2/WebGL and it needs to work across all current 3D APIs, so it is a bit too radical on the ‘simplicity side’ and too thin on the ‘feature side’, but it demonstrates the basic idea of a simple-yet-modern 3D API well I think.





About List